Atlas

Models

← All models

Ring 2.6 1T

InclusionAI · 2026-05-14 · 1T parameters

Ring 2.6 1T is InclusionAI's open-weight trillion-scale reasoning MoE model, published May 14, 2026 under the MIT license, and the reasoning companion to Ling 2.6 1T. It exposes high and xhigh reasoning-effort settings for agentic coding, tool use, search, and long-horizon workflows, and uses a 7:1 hybrid of Lightning Attention and MLA with a documented maximum context of 262,144 tokens. The released checkpoint contains exactly 1,025,657,871,744 parameters, while InclusionAI describes its activated size only approximately as 63B.

Benchmark scores

BenchmarkScore
AA-LCR64.3
AA-Omniscience Index-37.8
Artificial Analysis Agentic Index18.9
Artificial Analysis Coding Index42.8
Artificial Analysis Intelligence Index30.6
Artificial Analysis Omniscience Accuracy25.1
Artificial Analysis Omniscience Hallucination Rate83.8
Artificial Analysis Openness Index38.9
AutomationBench-AA2.1
BLXBench65.3
BLXBench Coding category96.8
BLXBench Cost category84.8
BLXBench Debugging category50.8
BLXBench Hallucination category74.1
BLXBench pass rate44.7
BLXBench Reasoning category69.6
BLXBench Refactoring category54.7
BLXBench Security category54.7
BLXBench Speed category84.2
BLXBench UI category35.4
BuseyBench SVG3.4
Capability148.0
CRED v0.9267.5
CRED v0.92 - Data integrity92.5
CRED v0.92 - Execution100.0
CRED v0.92 - Paper–code41.1
CRED v0.92 - Prose–table82.0
CritPt3.7
GDPval-AA v2920
GDPval-AA v2 (normalized)21.0
Loading Atlas data…